Coevolution-informed Bayesian optimization for sample-efficient protein design

Protein engineering is limited less by generating variants than by the cost of evaluating them, so designing under a tight budget demands sequence features that let a model learn fitness from very few examples. We introduce ALSEBO (Active Learning Sequence Exploration via Bayesian Optimization), which couples a generative latent sequence landscape to Bayesian optimization and featurizes candidates with direct-coupling-analysis (DCA) coevolutionary statistics. This representation carries a specific inductive bias: it places the dominant organizer of the fitness landscape along a single linear coordinate, producing a smooth, funnel-like objective that a low-data surrogate navigates efficiently. On a virtual avGFP fluorescence benchmark, ALSEBO reaches the optimum in 40 evaluations and outpaces protein-language-model embeddings and raw latent coordinates; controls with representation-neutral oracles confirm that the advantage is intrinsic, not an artifact of the benchmark. Molecular dynamics of the optimized variant recovers structural hallmarks of fluorescence, and ALSEBO transfers to divergent GFP orthologs and to a non-GFP enzyme, establishing a data-efficient route to protein design.